深度学习实践——图像分类：训练数据增广

图像增广(augmentation)是在有限训练数据集的情况下，有效扩大数据规模的方式。它的作用有：
1）通过随机增广由一张图片产生一组互不相同的图片，扩大数据规模
2）通过位移/旋转/缩放等方式的增广，能够减小模型对于图像中物体位置的敏感度
3）通过调节颜色/亮度等方式，减小对相应属性的敏感度
4）模糊/噪点/降低分辨率，减小对边界及图片细节像素的敏感度

训练数据增广办法：

1. 位移：

import numpy as np
import cv2

def shift(img, delta_width, delta_height):
    height,width,channel = img.shape
    mat = np.float32([[1,0,delta_width],[0,1,delta_height]])
    new_img = cv2.warpAffine(img, mat, (width,height))
    
    return new_img

比较处理前后变化

import matplotlib.pyplot as plt
import cv2
img_file = "~/dog.jpeg"
img0 = cv2.cvtColor(cv2.imread(img_file), cv2.COLOR_BGR2RGB)
img = shift(img0, 50, 20)

image.png

2. 缩放

def zoom(img, scale):
    new_img = cv2.resize(img, (0,0), fx=scale, fy=scale)
    if scale <= 1:
        top = bottom = int((img.shape[0]-new_img.shape[0])/2)
        left = right = int((img.shape[1]-new_img.shape[1])/2)
        new_img = cv2.copyMakeBorder(new_img.copy(), top, bottom, left, right,cv2.BORDER_CONSTANT, value=[0, 0, 0])
    else:
        top = bottom = int((new_img.shape[0]-img.shape[0])/2)
        left = right = int((new_img.shape[1]-img.shape[1])/2)
        new_img = new_img[top:-top,left:-left,:]
    
    return new_img

image.png

3. 旋转

def rotate(img, aug_value):
    M = cv2.getRotationMatrix2D((0.5*img.shape[1], 0.5*img.shape[0]), aug_value, 1.0)
    img = cv2.warpAffine(img.astype(np.uint8), M, (img.shape[1], img.shape[0]), flags=cv2.INTER_LINEAR)
    return img

image.png

4. 翻转

img1 = np.fliplr(img)  # 水平翻转
img2 = np.flipud(img) # 上下翻转

image.png

5. 颜色变化

6. 亮度变化

7. 对比度变化

#直方图均衡，可用于调整图片的局部对比度
for ch in range(img.shape[2]):
    img[..., ch] = cv2.equalizeHist(img[..., ch])

image.png

8. 模糊

#高斯模糊
def blur(img, value):
    return cv2.GaussianBlur(img0, (value, value), 0)

image.png

9. 降低分辨率

def zoom_blur(img,aug_value):
    img0 = cv2.resize(img, (0, 0), fx=aug_value,fy=aug_value)
    img0 = cv2.resize(img0, (0, 0), fx=1/aug_value,fy=1/aug_value, interpolation=cv2.INTER_NEAREST)
    return img0

image.png

10. 增加噪点

def gaussian_noise(img, aug_value):
    noise = np.zeros(img.shape)
    num_ch = 1 if img.ndim == 2 else img.shape[2]
    cv2.randn(noise, tuple([0]*num_ch), tuple([aug_value]*num_ch))
    img0 = np.clip(img.astype(np.float)+noise, 0, 255).astype(np.uint8)
    return img0

image.png